Advances in Aerodynamic Shape Optimization

Antony Jameson
Stanford University, Stanford, CA 94305-4035 jameson@baboon.stanford.edu

1 Introduction
The focus of CFD applications has shifted to aerodynamic design. This shift has been mainly motivated by the availability of high performance computing platforms and by the development of new and ecient analysis and design algorithms. In particular automatic design procedures, which use CFD combined with gradient-based optimization techniques, have had a signicant impact on the design process by removing diculties in the decision making process faced by the aerodynamicist. A fast way of calculating the accurate gradient information is essential since the gradient calculation can be the most time consuming portion of the design algorithm. The computational cost of gradient calculation can be dramatically reduced by the control theory approach since the computational expense incurred in the calculation of the complete gradient is eectively independent of the number of design variables. The foundation of control theory for systems governed by partial dierential equations was laid by J.L. Lions [1]. The method was rst used for aerodynamic design by Jameson in 1988 [2, 3]. Since then, the method has even been successfully used for the aerodynamic design of complete aircraft congurations [4]. In the present work a continuous adjoint formulation has been used to derive the adjoint system of equations, in which the adjoint equations are derived directly from the governing equations and then discretized. This approach has the advantage over the discrete adjoint formulation in that the resulting adjoint equations are independent of the form of discretized ow equations. The adjoint system of equations has a similar form to the governing equations of the ow, and hence the numerical methods developed for the ow equations [5, 6, 7] can be reused for the adjoint equations. Moreover, the gradient can be derived directly from the adjoint solution and the surface motion, independent of the mesh modication. In order to accelerate the convergence of the descent process the gradient is then smoothed implicitly via a second order dierential equation. This is equivalent to redening the gradient in a Sobolve space. The resulting procedure is very ecient, often yielding the optimum in 10-20 design cycles. Recently wing planform parameters have been included as design variables and the Aerospace Computing Laboratory at Stanford University has successfully designed a wing which produces a specied lift with minimum

688

Antony Jameson

drag, while meeting other criteria such as low structure weight, sucient fuel volume, and stability and control [8]. Based on the promising results from our wing planform optimization strategy applied to inviscid ow and from our viscous aerodynamic design techniques [9, 10], we are now applying wing shape and planform optimization methods to viscous ow in order to take into account the viscous eects such as shock/boundary layer interaction, ow separation, and skin friction and eventually produce more realistic designs [11]. Additionally, the design method, which is greatly accelerated by the use of control theory, has been further enhanced by the use of a new continuous adjoint method, which reduce the volume integral part of the adjoint gradient formula to a surface integral [12], thus eliminating the dependence of the gradient formulas on the mesh perturbation. The computational savings in the gradient calculation are particularly signicant for three-dimensional aerodynamic shape optimization problems on general unstructured and overset meshes. The use of unstructured grid techniques hold considerable promise for aerodynamic design by facilitating the treatment of complex congurations without incurring a prohibitive cost and bottleneck in mesh generation. The computational feasibility of using unstructured meshes for design is essentially enabled by the use of the continuous adjoint approach and the reduced gradient formulas [13]. Representative calculations are displayed in gures 2 through 6.

2 Adjoint and Gradient Formulations for the Equations of Transonic Flow

The adjoint method may be applied directly to the partial dierential equations to derive a continuous adjoint equation, which must then be discretized to obtain a numerical solution. Alternatively one may derive a discrete adjoint equation directly after rst discretizing the ow equations. In this work the rst procedure has been adopted because it allows more exibility in the formulation of the gradient. The procedure is illustrated here for the Euler equations. These are represented in transformed coordinates i on a xed computational domain. Let S = JK 1 where Kij = xi , J = det(K) j

Then the transformed equations are Fi (Sij fj ) = =0 i i

Advances in Aerodynamic Shape Optimization

689

As an example, consider the case of an inverse problem where one wishes to nd the shape which brings the pressure as close as possible to the specied target pressure, pt . Hence we try to minimize the cost function I= 1 2 (p pt )2 dS

over the design surface B, which for convenience is assumed to be the surface 2 = 0. Now a shape modication induces a change p in the pressure and consequently 1 (p pt )2 dS I = (p pt )pdS + 2 B B Also the change in the solution is given by (Fi (w)) = 0 i Here the ux changes are Fi = Sij fj + Ci w where fj w Consequently one can augment the cost variation by Ci = Sij T
D

Fi d = i

ni T Fi dB

T Fi d

Now choose to satisfy the adjoint equation

T Ci

=0 i

with the boundary condition 2 x + 3 y + 4 z = p pt where x , y , z are the components of the surface normal. Then the boundary integrals involving p and the eld integral involving w are eliminated and the gradient is reduced to 1 2 (p pt )2 dS (S21 2 + S22 3 + S23 4 ) pd1 d3 T (Sij fj )d

where typically the rst term is negligible and can be dropped. Other cost functions, such as the drag coecient, lead to dierent adjoint boundary

690

Antony Jameson

conditions. When a mesh perturbation procedure is dened the perturbations Sij can be directly related to the surface perturbation. Finally one obtains the cost variation as an inner product I = (G, F ) = GF d1 d2

where G is the pointwise gradient, and F denes the surface shape.

3 Optimization Procedure
Another key issue for successful implementation of the continuous adjoint method is the choice of an appropriate inner product for the denition of the gradient. It turns out that there is an enormous benet from the use of a modied Sobolev gradient, which enables the generation of a sequence of smooth shapes. The gradient G is generally in a lower smoothness class than the initial shape F . Then a sequence of steps F = G progressively reduces the smoothness, leading to instability. In order to prevent this we can introduce a weighted Sobolev inner product [14] of the form u, v = (uv + u v )dx

in one dimension, where the parameter controls the smoothness. Corre spondingly we dene a modied gradient G such that I =< G, F > . In the one dimensional case G is obtained by solving the smoothing equation G = G. G 1 1 (1)

In the multi-dimensional case the smoothing is applied in product form. Finally we set F = G (2) with the result that I = < G, G > < 0, unless G = 0, and correspondingly G = 0. The implicit smoothing procedure acts as a preconditioner which allows the use of much larger steps for the search procedure and leads to a large reduction in the number of design iterations needed for convergence.

Advances in Aerodynamic Shape Optimization

691

Flow Solution

Adjoint Solution

Gradient Calculation Repeat the Design Cycle until Convergence

Sobolev Gradient

Shape & Grid Modification

Fig. 1. Design cycle

3.1 Outline of the design procedure The design procedure can nally be summarized as follows: 1. Solve the ow equations for , u1 , u2 , u3 , p. 2. Solve the adjoint equations for subject to appropriate boundary conditions. 3. Evaluate G and calculate the corresponding Sobolev gradient G. into an allowable subspace that satises any geometric con4. Project G straints. 5. Update the shape based on the direction of steepest descent. 6. Return to 1 until convergence is reached. Practical implementation of the design method relies heavily upon fast and accurate solvers for both the state (w) and co-state () systems. The result obtained in Section 4 have been obtained using well-validated software for the solution of the Euler and Navier-Stokes equations developed over the course of many years [5, 15, 16]. For inverse design the lift is xed by the target pressure. In drag minimization it is also appropriate to x the lift coecient, because the induced drag is a major fraction of the total drag, and this could be reduced simply by reducing the lift. Therefore the angle of attack is adjusted during each ow solution to force a specied lift coecient to be attained, and the inuence of variations of the angle of attack is included in the calculation of the gradient. The vortex drag also depends on the span loading, which may be constrained by other considerations such as structural loading or buet onset. Consequently, the option is provided to force the span loading by adjusting the twist distribution as well as the angle of attack during the ow solution. The design procedure has been implemented in the computer programs Syn88 and Syn107 for three-dimensional wing-fuselage design using the Euler

692

Antony Jameson

and RANS equations respectively on structured meshes. Both codes include automatic mesh generation. The method has also been implemented for unstructured meshes in the computer program Synplane, which treats complete aircraft.

4 Case Studies
4.1 B747 Euler planform result The shape changes in the section needed to improve the transonic wing design are quite small. However, in order to obtain a true optimum design larger scale changes such as changes in the wing planform (sweepback, span, chord, section thickness, and taper) should be considered. Because these directly aect the structure weight, a meaningful result can only be obtained by considering a cost function that accounts for both the aerodynamic characteristics and the weight. In references [8, 11, 17] the cost function is dened as I = 1 CD + 2 1 2 (p pd )2 dS + 3 CW ,

where CW qWref is a dimensionless measure of the wing weight, which S can be estimated either from statistical formulas, or from a simple analysis of a representative structure, allowing for failure modes such as panel buckling. The coecient 2 is introduced to provide the designer some control over the pressure distribution, while the relative importance of drag and weight are represented by the coecients 1 and 3 . By varying these it is possible to calculate the Pareto front of designs which have the least weight for a given drag coecient, or the least drag coecient for a given weight. The relative importance of these can be estimated from the Breguet range equation; R = R = CD 1 W2 + W1 CD log W2 W2 CW CD 1 + W1 W CD log W2 q S2 ref .

Figure 2 shows the Pareto front obtained from a study of the Boeing 747 wing [17], in which the ow was modeled by the Euler equations. The wing planform and section were varied simultaneously, with the planform dened by six parameters; sweepback, span, the chord at three span stations, and wing thickness. The weight was estimated from an analysis of the section thickness required in the structural box. The gure also shows the point on the Pareto front when 3 is chosen such that the range of the aircraft is 1

Advances in Aerodynamic Shape Optimization

693

Pareto front 0.052

0.05 optimized section with fixed planform 0.048 baseline

maximized range

0.046
C

w 0.044 0.042 = optimized section and planform 0.04 0.038 80

85

90
C

95 D (counts)

100

105

110

Fig. 2. Pareto front of section and planform modications

maximized. The optimum wing, as illustrated in gure 3, has a larger span, a lower sweep angle, and a thicker wing section in the inboard part of the wing. The increase in span leads to a reduction in the induced drag, while the section shape changes keep the shock drag low. At the same time the lower sweep angle and thicker wing section reduce the structural weight. Overall, the optimum wing improves both aerodynamic performance and structural weight. The drag coecient is reduced from 108 counts to 87 counts (19%), while the weight factor CW is reduced from 455 counts to 450 counts (1%).

Fig. 3. Superposition of the baseline (green) and the optimized section-andplanform (blue) geometries of Boeing 747. The redesigned geometry has a longer span, a lower sweep angle, and thicker wing sections, improving both aerodynamic and structural performances. The optimization is performed at Mach .87 and xed CL .42, where 3 is chosen to maximize the range of the aircraft. 1

694

Antony Jameson
Cp = -2.0 Cp = -2.0

B747 WING-BODY
Mach: 0.780 Alpha: 2.683 CL: 0.449 CD: 0.01137 CM:-0.1369 Design: 30 Residual: 0.1710E-02 Grid: 257X 65X 49

B747 WING-BODY
Mach: 0.850 Alpha: 2.220 CL: 0.449 CD: 0.01190 CM:-0.1498 Design: 30 Residual: 0.7857E-03 Grid: 257X 65X 49

Tip Section: 92.5% Semi-Span Cl: 0.453 Cd:-0.01561 Cm:-0.2117 Cp = -2.0 Cp = -2.0 Cp = -2.0

Tip Section: 92.5% Semi-Span Cl: 0.462 Cd:-0.01878 Cm:-0.2213 Cp = -2.0

Root Section: 13.0% Semi-Span Cl: 0.344 Cd: 0.05089 Cm:-0.1171

Mid Section: 50.6% Semi-Span Cl: 0.569 Cd: 0.00036 Cm:-0.2516

Root Section: 13.0% Semi-Span Cl: 0.335 Cd: 0.05928 Cm:-0.1213

Mid Section: 50.6% Semi-Span Cl: 0.572 Cd:-0.00217 Cm:-0.2602

(a) Mach .78

Cp = -2.0

(b) Mach .85

B747 WING-BODY
Mach: 0.870 Alpha: 1.997 CL: 0.449 CD: 0.01224 CM:-0.1590 Design: 30 Residual: 0.3222E-03 Grid: 257X 65X 49

Drag rise at fixed CL

280
Tip Section: 92.5% Semi-Span Cl: 0.464 Cd:-0.02110 Cm:-0.2222 Cp = -2.0 Cp = -2.0

260 Baseline Redesign 240

220

CD (counts)

200

180

160

140

120

Root Section: 13.0% Semi-Span Cl: 0.332 Cd: 0.06246 Cm:-0.1273

Mid Section: 50.6% Semi-Span Cl: 0.574 Cd:-0.00334 Cm:-0.2674

100 0.78

0.8

0.82

0.84

0.86

0.88

0.9

0.92

Mach

(c) Mach .87

(d) Drag Vs. Mach number

Fig. 4. (a)-(c): Super B747 at Mach .78, .85, and .87 respectively. Dash line represents shape and pressure distribution of the initial conguration. Solid line represents those of the redesigned conguration. (d): Drag Vs. Mach number of Super B747.

Advances in Aerodynamic Shape Optimization Table 1. Comparison between Baseline B747 and Super B747 at Mach .86 CL Boeing 747 Super B747 .45 .50 CD CW counts counts 141.3 499 (107.0 pressure, 34.3 viscous) (82,550 lbs) 141.9 427 (104.8 pressure, 37.1 viscous) (70,620 lbs)

695

4.2 Super B747 In order to explore the limits of attainable performance the B747 wing has been replaced by a completely new wing to produce a Super B747. An initial design was created by blending supercritical wing sections obtained from other optimizations to the optimum planform which was found in the planform study described in the previous section. Then the RANS optimization code Syn107 was used to obtain minimize drag over 3 design points at Mach .78, .85, and .87, shown in gures 4 (a)-(c) with a xed lift coecient of .45 for the exposed wing, corresponding to a lift coecient of about .52 when the fuselage lift is included. Because the new wing sections are signicantly thicker, the new wing is estimated to be 12,000 pounds lighter than the baseline B747 wing as shown in table 1. At the same time the drag is reduced over the entire range from Mach .78 to .90 with a maximum benet of 25 counts at Mach .87, as shown in gure 4 (d). Figure 5 and table 2 display the lift-drag polar at Mach .86. The drag coecient of the Super B747 is 142 counts at a lift coecient of .5, whereas the baseline B747 has the same drag at a lift coecient of .45. This represents improvement in L/D of more than
Table 2. Comparison of drag polar; B747 Vs. Super B747 Boeing 747 Super B747 CL CD CL CD 0.0045 94.3970 0.0009 76.9489 0.0505 67.8010 0.0500 82.2739 0.1000 74.6195 0.1005 64.6147 0.1501 72.1087 0.1506 65.5073 0.2006 69.4840 0.2002 73.9661 0.2503 79.6424 0.2507 76.0041 0.3005 88.7551 0.3008 84.9889 0.3507 101.5293 0.3509 95.6117 0.4010 106.9625 0.4009 118.0487 0.4512 141.2927 0.4510 121.7183 0.5014 177.0959 0.5010 141.8675 0.5512 175.2569 0.5516 228.1786 0.6016 298.0458 0.6014 222.5459 (CD in counts) Note equal drag of the baseline B747 at CL .45 and the Super B747 at CL .5.

696

Antony Jameson

Fig. 5. Drag Polars of Baseline and Super B747 at Mach .86. (Solid-line represents Super B747. Dash-line represents Baseline B747.)

10 percent. In combination with the reduction in wing weight and an increase in fuel volume due to the thicker wing section, this should lead to an increase in range which is substantially more than 10 percent. 4.3 Shape optimization for a transonic business Jet The unstructured design method has also been applied to several complete aircraft congurations. The results for a business jet are shown in gures 6 (a) and (b). There is a strong shock over the out board wing sections of the initial conguration, which is essentially eliminated by the redesign. The drag was reduced from 235 counts to 215 counts in about 8 design cycles. The lift was constrained at 0.4 by perturbing the angle of attack. Further, the original thickness of the wing was maintained during the design process

Advances in Aerodynamic Shape Optimization

697

(a) Baseline

(b) Redesign

Fig. 6. Density contours for a business jet at M = 0.8, = 2o

ensuring that fuel volume and structural integrity will be maintained by the redesigned shape. Thickness constraints on the wing were imposed on cutting planes along the span of the wing and by transferring the constrained shape movement back to the nodes of the surface triangulation.

5 Conclusion
The accumulated experience of the last decade suggests that most existing aircraft which cruise at transonic speeds are amenable to a drag reduction of the order of 3 to 5 percent, or an increase in the drag rise Mach number of at least .02. These improvements can be achieved by very small shape modications, which are too subtle to allow their determination by trial and error methods. When larger scale modications such as planform variations or new wing sections are allowed, larger gains in the range of 5-10 percent are attainable. The potential economic benets are substantial, considering the fuel costs of the entire airline eet. Moreover, if one were to take full advantage of the increase in the lift to drag ratio during the design process, a smaller aircraft could be designed to perform the same task, with consequent further cost reductions. Methods of this type will provide a basis for aerodynamic designs of the future.

References
1. J. L. Lions. Optimal Control of Systems Governed by Partial Dierential Equations. Springer-Verlag, New York, 1971. Translated by S.K. Mitter.

698

Antony Jameson

2. A. Jameson. Aerodynamic design via control theory. Journal of Scientic Computing, 3:233260, 1988. 3. A. Jameson. Computational aerodynamics for aircraft design. Science, 245:361 371, July 1989. 4. J. Reuther, A. Jameson, J. Farmer, L. Martinelli, and D. Saunders. Aerodynamic shape optimization of complex aircraft congurations via an adjoint formulation. AIAA paper 96-0094, 34th Aerospace Sciences Meeting and Exhibit, Reno, Nevada, January 1996. 5. A. Jameson, W. Schmidt, and E. Turkel. Numerical solutions of the Euler equations by nite volume methods with Runge-Kutta time stepping schemes. AIAA paper 81-1259, January 1981. 6. A. Jameson and T.J. Baker. Improvements to the aircraft Euler method. AIAA paper 87-0452, AIAA 25th Aerospace Sciences Meeting, Reno, Nevada, January 1987. 7. T. J. Barth. Apects of unstructured grids and nite volume solvers for the Euler and Navier Stokes equations. AIAA paper 91-0237, AIAA Aerospace Sciences Meeting, Reno, NV, January 1991. 8. K. Leoviriyakit and A. Jameson. Aerodynamic shape optimization of wings including planform variations. AIAA paper 2003-0210, 41st Aerospace Sciences Meeting & Exhibit, Reno, Nevada, January 2003. 9. A. Jameson. A perspective on computational algorithms for aerodynamic analysis and design. Progress in Aerospace Sciences, 37:197243, 2001. 10. S. Kim, J. J. Alonso, and A. Jameson. Design optimization of high-lift congurations using a viscous continuous adjoint method. AIAA paper 2002-0844, AIAA 40th Aerospace Sciences Meeting & Exhibit, Reno, NV, January 2002. 11. K. Leoviriyakit, S. Kim, and A. Jameson. Viscous aerodynamic shape optimization of wings including planform variables. AIAA paper 2003-3498, 21st Applied Aerodynamics Conference, Orlando, Florida, June 23-26 2003. 12. A. Jameson and S. Kim. Reduction of the adjoint gradient formula in the continuous limit. AIAA paper 2003-0844, AIAA 41st Aerospace Sciences Meeting & Exhibit, Reno, NV, January 2003. 13. A. Jameson, Sriram, and L. Martinelli. An unstructured adjoint method for transonic ows. AIAA paper, 16th AIAA CFD Conference, Orlando, FL, June 2003. 14. A. Jameson, L. Martinelli, and J. Vassberg. Reduction of the adjoint gradient formula in the continuous limit. AIAA paper, 41st AIAA Aerospace Sciences Meeting, Reno, NV, January 2003. 15. L. Martinelli and A. Jameson. Validation of a multigrid method for the Reynolds averaged equations. AIAA paper 88-0414, 1988. 16. S. Tatsumi, L. Martinelli, and A. Jameson. A new high resolution scheme for compressible viscous ows with shocks. AIAA paper To Appear, AIAA 33nd Aerospace Sciences Meeting, Reno, Nevada, January 1995. 17. K. Leoviriyakit and A. Jameson. Aero-structural wing planform optimization. Reno, Nevada, January 2004. Proceedings of the 42st Aerospace Sciences Meeting & Exhibit.